Fix CP norm for complex weights - #656

cp_norm and CPTensor.norm() can return an imaginary value for a tensor with complex CP weights. For a weight of 1j and real factors [[1], [2]] and [[3], [4]], both return 11.180339887498949j, while the represented tensor has Frobenius norm 11.180339887498949.

Conjugate the second weight axis in the weight outer product, extending the complex-factor support introduced by @aarmey in #423 to complex weights. The squared norm is real, so take its real part before the square root. This also supports Paddle CPU, whose complex sqrt kernel is unavailable.

Expose the backend real operation through the existing dispatch mechanism. Register native operations for the dense backends and NumPy sparse backend. Paddle returns an already-real input unchanged because its native real kernel accepts complex tensors only. Add scalar/vector real-part tests for float32/float64/complex64/complex128, norm output dtype checks, and float32/float64 weighted and unweighted controls.

Local validation on Python 3.12.14, with the core tensor algebra:

• Windows, NumPy 2.3.5 / SciPy 1.16.3 / Sparse 0.19.2: complete suite 791 passed, 3 skipped, 1 xfailed. Eight additional public sparse real-part controls preserve dtype, shape, sparse structure and inputs.
• Windows, PyTorch 2.8.0 on CPU: complete suite 786 passed, 4 skipped, 1 xfailed, with NumPy and Torch random seeds set to 1.
• Ubuntu through WSL, Paddle 3.0.0b1 on CPU: complete suite 702 passed, 4 skipped, 1 xfailed. The new real operation covers real and complex input dtypes. This run is bound to the 12 modified core files; the later sparse-only registration is covered by the NumPy complete suite and the eight sparse controls above.
• Keeping the old weight formula while sharing the new tests, real wrappers and real-valued norm projection gives 8 failures and 8 controls for the 16 complex-weight cases on Paddle; all 16 pass with the corrected formula. This is a controlled comparison over the exact base archive, with those explicit overlays.
• Black 26.10.0's Python API checks all 170 tracked Python files, using the project target versions and fast=False safety checks; every file is unchanged. The CLI traversal attempts did not complete.

These validation ranges overlap. JAX, TensorFlow and CuPy bindings were inspected statically; their runtime suites were not run locally. Existing Windows test dependencies were reused, and the isolated Paddle environment uses Python 3.12.14 while official CI currently uses 3.12.15. GPU, model accuracy and performance tests were not run. Test tolerances, skip conditions and CI configuration are unchanged.

Fixes #655.